Designing a pre-execution governance gate for AI calls.
Most governance systems observe a decision after it happened. We are testing a stricter boundary: a deterministic governance decision must complete before an external AI request is released.
Logging after a model call cannot stop that call.
Article 12 of the EU AI Act addresses automatic logging capabilities for high-risk AI systems. Article 14 addresses effective human oversight. Neither article prescribes one database or consensus algorithm. Read the regulation ↗
For an engineering team, however, the operational question is unavoidable: where is the control decision made, and can the system prove which policy, evidence and accountable actor produced it?
Putting every request behind a cross-region consensus round can make the governance layer the slowest component. Avoiding the check is not acceptable. Pretending that the network is free is not engineering.
Local decision. Serializable control record.
Classify the request in its home region.
Load a versioned policy bundle and current control state locally.
Return allow, deny or require_confirmation deterministically.
Append the decision envelope to an ordered control log.
Release the external model request only after allow.
Reconcile evidence and confirmations into the audit state.
The topology is SLOG-inspired. The current production repository is not presented as an implementation of the SLOG paper. That name remains a design reference until topology and benchmark evidence are published. Read the PVLDB paper ↗
Versioned evidence with explicit predecessor links.
The current product contains versioned evidence records, SHA-256 verification, predecessor links, snapshot-chain construction, recurring confirmation state and audit exports.
{
"evidenceVersionId": "ev_0187",
"sha256": "9d9f...3a2c",
"prevVersionId": "ev_0186",
"recordedAt": "2026-07-27T09:12:00Z",
"controlId": "human-oversight"
}snapshotHash[n] =
SHA-256(
canonical(snapshot[n])
|| snapshotHash[n-1]
)A SHA-256-linked record is not automatically immutable. Immutability also depends on authorization, storage controls, retention, key custody and independent anchoring. Ed25519 signatures are a planned hardening step, not a current production claim.
The benchmark must travel with the claim.
The model call is never released after a deny, timeout or unavailable policy snapshot.
Identical canonical inputs and policy versions produce the same decision envelope.
Concurrent decisions receive a reproducible order in the authoritative control log.
p50, p95 and p99 overhead are measured by region, payload size, cache state and failure mode.
The latency target is an experiment, not a headline. Numbers will be published only with the harness, environment, sample size and raw result artifact.
Connect oversight and evidence to the lifecycle.
The EU AI Act does not say to use this architecture. It creates a need to connect logging, oversight, responsibilities and evidence to the actual system lifecycle.
- the policy version becomes part of the decision record.
- a required human confirmation can block release.
- evidence lineage can be inspected independently.
- audit exports can reproduce the control state at a point in time.
Challenge the boundary.
- Which failure mode breaks the ordering assumption?
- Where should an independent integrity anchor sit?
- Which benchmark distribution would make the latency result credible?
- What would you require before calling the record tamper-evident?